Key Insight: Evidence Must Drive AI Policy

The authors stress that the idea of evidence itself is not simple or straightforward. What qualifies as evidence can vary across fields: in health policy, randomized control trials serve as the gold standard, while in economics, forecasts and theoretical models hold weight. And history illustrates that evidence will often be questioned or ignored: the tobacco industry leaned on inconclusive studies to stall public health measures, and fossil fuel companies downplay climate risks despite knowing otherwise. These examples show that the tasks of defining, evaluating, and acting on evidence are urgent and complex. In response to these challenges, the authors encourage the US government to utilize the Foundations for Evidence-Based Policymaking Act (Evidence Act).

Key Insight: Policy Can Accelerate Evidence Generation

The authors propose several mechanisms to make policy the driver of evidence creation. Policymakers should incentivize pre-release evaluations, ensuring that risks (such as using AI for malicious purposes, the likelihood of AI hallucinations, or the prevalence of AI generating copyrighted material) are measured before companies deploy new models. They also call for increased transparency, citing findings from the 2024 Foundation Model Transparency Index that top AI companies fall short when it comes to publicly reporting their risk-mitigation practices. They recommend post-deployment monitoring, such as adverse-event reporting systems that track concrete instances of harm once models are in use. Finally, they encourage protections for third-party research, noting that independent investigators often face legal and contractual barriers when probing AI systems. Safe harbor provisions, modeled on cybersecurity law, would enable such research to proceed in the public interest. Together, these measures would expand the evidence base and allow AI policy to evolve in step with the technology itself.

Key Insight: Consensus in a Fragmented Field

The AI research and policy community is currently divided, with divergent views on the seriousness of risks and the speed of technological progress. This lack of alignment makes it difficult to establish clear, effective policy responses. Drawing from precedents in climate governance and disaster policy, the authors call for deliberate processes that foster consensus, even amid uncertainty. Global initiatives, such as the UN鈥檚 High-Level Advisory Board on AI and proposals for an International Scientific Panel, aim to provide shared baselines of evidence. Such consensus would not eliminate debate but would ensure that disagreements unfold with a common evidentiary framework, strengthening the legitimacy and durability of policy decisions.

Why This Matters

As AI becomes more central to business operations, having trustworthy and reliable systems will be crucial. Business leaders and executives will benefit from understanding the growing landscape of AI policy, supporting evidence-based foundations for AI technology, and following the guidance of institutions that produce independent research. By aligning with these principles, companies will not only be ready to comply with emerging regulations, but will also be a step ahead to build trust with customers and stakeholders. As the authors conclude, governing AI will be one of the grand challenges of the 21st century, and informed business leaders will have an important role to play facing it.

References

[1] Rishi Bommasani et al., “Advancing science- and evidence-based AI policy,鈥� Science 389 (2025): 459. DOI:

[2] Bommasani et al., “Advancing science- and evidence-based AI policy,鈥� 460.

[3] Bommasani et al., “Advancing science- and evidence-based AI policy,鈥� 461.

Meet the Authors

is an Assistant Professor of Business Administration at 性视界 Business School and PI in D^3鈥檚 Trustworthy AI Lab. She is also a faculty affiliate in the Department of Computer Science at 性视界 University, the 性视界 Data Science Initiative, Center for Research on Computation and Society, and the Laboratory of Innovation Science at 性视界. Professor Lakkaraju’s research focuses on the algorithmic, practical, and ethical implications of deploying AI models in domains involving high-stakes decisions such as healthcare, business, and policy.

Additional Authors: Rishi Bommasani, Sanjeev Arora, Jennifer Chayes, Yejin Choi, Mariano-Florentino Cu茅llar, Li Fei-Fei, Daniel E. Ho, Dan Jurafsky, Sanmi Koyejo, Arvind Narayanan, Alondra Nelson, Emma Pierson, Joelle Pineau, Scott Singer, Ga毛l Varoquaux, Suresh Venkatasubramanian, Ion Stoica, Percy Liang, and Dawn Song

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post Evidence at the Core: How Policy Can Shape AI鈥檚 Future appeared first on 性视界 Business School AI Institute.

Key Insight: Accelerate or Anchor?

The study identifies a foundational behavioral pattern: when an agent鈥檚 current task closely resembles a stored memory, the outputs tend to closely match as well. This 鈥渆xperience-following鈥� correlation mirrors how humans often rely on familiar patterns, and it can accelerate learning when the stored example is correct. However, it鈥檚 also not without risks. If erroneous or low-quality experiences are stored in memory, they can be applied to future tasks, thereby decreasing the agent鈥檚 overall performance. This means that the quality of stored examples is paramount, as bad memories don鈥檛 just linger, they can create a propagating error feedback loop.

Key Insight: Selective Addition

If the experience-following property shows why quality matters in LLM agents, then addition shows how to control it, and a clear finding from the study is that indiscriminate memory growth actually hurts performance. In tests with three different agents, covering electronic health records (EHRs), the LLM-based autonomous driving agent AgentDriver, and a network security agent, storing every task and output (鈥渁dd-all鈥�) performed worse than using no memory addition at all. However, using strict evaluation criteria and filtering before storage led to an average 10% performance boost, so memory improvement is less about hoarding information than curating a high-quality knowledge base.

Key Insight: Improvement through Deletion

Even with careful addition, not all stored experiences are equally useful over time. Some look similar to new tasks (鈥渉igh input similarity鈥�), but consistently produce poor output (鈥渓ow output similarity鈥�). The authors term this 鈥渕isaligned experience replay,鈥� and show that pruning these entries improves long-term outcomes. Removing experiences with repeatedly low utility (鈥渉istory-based deletion鈥�) offered the best boost to performance while effectively and efficiently maintaining memory size. From a strategic perspective, this practice mirrors audits of playbooks, datasets, and best practices to ensure that institutional knowledge remains in top shape.

Why This Matters

The results from this research should give business leaders important context for thinking about how to choose and deploy AI agents: more data isn鈥檛 automatically better, and AI鈥檚 鈥渆xperience鈥� can actually be a liability, entrenching errors and bloating infrastructure. Disciplined curation, by selectively adding high-value experiences and strategically deleting low-value or misaligned ones, yields not only better accuracy but also more efficient, adaptable systems. In a world where executives may be involved in decision-making around LLM agents for their organizations, it鈥檚 important to have a blueprint for keeping AI agents sharp, reliable, and resilient, just like they plan for the training and advancement of their human employees. By understanding and investing in the processes that keep your AI鈥檚 memory in top-shape, your business will be equipped for tomorrow鈥檚 challenges.

References

[1] Zidi Xiong et al., 鈥淗ow Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior,鈥� arXiv preprint arXiv:2505.16067v1 (May 21, 2025): 2.

[2] Xiong et al., 鈥淗ow Memory Management Impacts LLM Agents,鈥� 5.

[3] Xiong et al., 鈥淗ow Memory Management Impacts LLM Agents,鈥� 9.

Meet the Authors

is a PhD student in computer science at 性视界 University, advised by Himabindu Lakkaraju.

is a PhD student in computer science at Michigan State University.

is a PhD student in computer science engineering at University of Minnesota – Twin Cities.

is a PhD student in computer science and engineering at Michigan State University.

is a University Foundation Professor in the computer science and engineering department at Michigan State University.

is an Assistant Professor of Business Administration at 性视界 Business School and PI in D^3鈥檚 Trustworthy AI Lab. She is also a faculty affiliate in the Department of Computer Science at 性视界 University, the 性视界 Data Science Initiative, Center for Research on Computation and Society, and the Laboratory of Innovation Science at 性视界. Professor Lakkaraju’s research focuses on the algorithmic, practical, and ethical implications of deploying AI models in domains involving high-stakes decisions such as healthcare, business, and policy.

is Assistant Professor of Computer Science at University of Georgia.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post Smarter Memories, Stronger Agents: How Selective Recall Boosts LLM Performance appeared first on 性视界 Business School AI Institute.

Key Insight: The Three Pillars of AI Trustworthiness

The holistic conceptualization taken by the authors in this paper recognizes that, for LLMs to be truly trustworthy, they must excel across multiple domains simultaneously. The researchers tested and demonstrated these principles using real-world datasets, including the Adult dataset, based on 1994 U.S. Census data, where they evaluated fairness by examining whether AI predictions of income varied based on gender attributes. Their privacy assessments used the Enron email dataset, containing over 600,000 emails with sensitive personal information including credit card numbers and Social Security Numbers. For robustness, they used the OOD Style Transfer, which incorporates text transformations, and AdvGLUE++ datasets, which includes examples for widely used Natural Language Processing (NLP) tasks.

Key Insight: Utilizing Novel Fine-Tuning Strategies

The 性视界 team developed two distinct strategies for embedding trustworthiness into AI systems. Their first approach, termed “Weak Trustworthiness Fine-tuning” (Weak TFT), focuses on training smaller models with explicit trustworthiness constraints, then using these models to teach larger systems. The second strategy, “Weak and Weak-to-Strong Trustworthiness Fine-tuning” (Weak+WTS TFT), applies trustworthiness constraints to both the small teacher model and the large student model during training.

Their experiments demonstrate that the Weak+WTS TFT approach produces significantly superior results, with improvements in fairness of up to 3 percentage points (equivalent to a 60% decrease in unfairness), as well as in robustness, or how resilient the AI was to attacks and unexpected situations. Remarkably, these ethical improvements required only minimal sacrifices in task performance鈥攄ecreases in accuracy did not exceed 1.5% across tested properties.

Key Insight: Challenges in Privacy Transfer

A key finding of the study is that not all trustworthiness properties transfer equally from weak to strong models. While the transfer of fairness and robustness properties showed promising results, privacy proved to be a more challenging attribute to transfer. The researchers found that larger models have a greater capacity to retain and recall details from their training data, which creates heightened vulnerabilities for exposing sensitive or confidential information. This finding highlights the complex nature of privacy in AI systems and suggests that different strategies may be needed to address privacy concerns in larger models.

Why This Matters:

For C-suite executives and business leaders, this research offers a potential pathway to developing more powerful LLM systems without compromising on certain ethical considerations. It suggests that companies could potentially start with smaller, more manageable models that are fine-tuned for trustworthiness in fairness and robustness, and then scale up to more capable systems while maintaining or even improving these critical properties. This approach could help mitigate risks associated with LLM deployment, enhance public trust in AI-driven decisions, and potentially reduce the resources required for ethical LLM development. However, the challenges identified in transferring privacy properties serve as a reminder of the complex nature of AI ethics. Business leaders should remain vigilant and consider multi-faceted approaches to ensuring the trustworthiness of their LLM systems, particularly when dealing with sensitive data.

Footnote

(1) The strong ceiling represents the benchmark performance of a large model that has been directly trained with trustworthiness constraints, serving as the upper bound for what the weak-to-strong approach should ideally achieve.

References

[1] Martin Pawelczyk et al., 鈥淕eneralizing Trust: Weak-to-Strong Trustworthiness in Language Models,鈥� arXiv preprint arXiv:2501.00418v1 (December 31, 2024): 1.

[2] Pawelczyk et al., 鈥淕eneralizing Trust,鈥� 2.

[3] Pawelczyk et al., 鈥淕eneralizing Trust,鈥� 8.

Meet the Authors

is a postdoctoral researcher at 性视界 working on trustworthy AI.

is an undergraduate student at 性视界 studying computer science.

is a PhD student in computer science at 性视界.

is a postdoctoral research associate at 性视界 working on trustworthy AI.

is an Assistant Professor of Business Administration at 性视界 Business School and PI in D^3鈥檚 Trustworthy AI Lab. She is also a faculty affiliate in the Department of Computer Science at 性视界 University, the 性视界 Data Science Initiative, Center for Research on Computation and Society, and the Laboratory of Innovation Science at 性视界. Professor Lakkaraju’s research focuses on the algorithmic, practical, and ethical implications of deploying AI models in domains involving high-stakes decisions such as healthcare, business, and policy.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post Teaching Trust: How Small AI Models Can Make Larger Systems More Reliable appeared first on 性视界 Business School AI Institute.

Key Insight: The Unified Attribution Framework

In this paper, Zhang and colleagues propose a unified framework that brings together three traditionally separate attribution methods: feature attribution (FA), which refers to the process of identifying which input features are most important in an AI model鈥檚 output, data attribution (DA), which involves understanding how specific training-data points influence an AI model鈥檚 behavior, and component attribution (CA), which focuses on understanding how internal parts of an AI model contribute to its output. This innovative approach recognizes that while these methods have evolved independently, they share fundamental techniques such as perturbations, gradients, and linear approximations. By unifying these methods, the researchers aim to provide a more comprehensive understanding of AI systems’ behavior.

Key Insight: Supporting Further Research

The unified framework offers multiple advantages for advancing AI interpretability research. For example, by promoting conceptual coherence through less fragmented terminology, it facilitates more effective communication and collaboration. The framework enables cross-attribution innovation, allowing researchers to adapt solutions developed for one attribution type to others, such as applying efficient sampling techniques from perturbation-based FA, which changes input parts to measure the effect on AI’s answers, to improve DA methods. It also simplifies theoretical analysis by identifying common mathematical underpinnings, streamlining research efforts and paving the way for more robust and generalizable techniques.

Key Insight: Implications for AI Regulation and Ethics

By providing a comprehensive view of AI system behavior, the unified attribution framework enables more informed and targeted regulatory approaches. The authors illustrate this with a real-world example: when tackling issues of bias in AI, the framework enables regulators to pinpoint potentially discriminating features in the input data, identify and track problematic or copyrighted training materials, and highlight specific components within the AI鈥檚 architecture that may contribute to biased outcomes. 

The authors note that regulation and policy frequently stress the need for transparency in AI systems and users’ right to an explanation. The unified attribution framework provides a powerful tool for practitioners to meet these legal and ethical requirements by offering detailed insights into both overall AI system behavior and specific input-output relationships.

Why This Matters

For business leaders, this unification method means gaining more comprehensive and reliable insights into how your AI systems function. Instead of fragmented views, leaders get a holistic understanding of what drives AI decisions. This is essential for building trust, ensuring regulatory compliance, and effectively identifying and addressing issues like bias or errors, whether they stem from data, inputs, or the model’s structure. Ultimately, the unified attribution framework proposed in this research supports more informed model management and governance, directly impacting an organization’s bottom line through cost savings and enhanced value.

References

[1] Shichang Zhang et al., 鈥淭owards Unified Attribution in Explainable AI, Data-Centric AI, and Mechanistic Interpretability,鈥� arXiv preprint arXiv:2501.18887v3 (May 29, 2025): 1.

[2] Zhang et al., 鈥淭owards Unified Attribution,鈥� 8.

[3] Zhang et al., 鈥淭owards Unified Attribution,鈥� 8.

Meet the Authors

is a postdoctoral fellow at the D^3 Institute at 性视界 University working with Professor Hima Lakkaraju. He received his Ph.D. in Computer Science from University of California, Los Angeles (UCLA).

is a PhD student in the Bioinformatics and Integrative Genomics Program at 性视界 Medical School.

is a PhD student in the 性视界 Computer Science program, working on machine learning interpretability and advised by Hima Lakkaraju. She is a strong advocate for increasing diversity in CS through direct mentorship of early-career minority students.

is an Assistant Professor of Business Administration at 性视界 Business School and PI in D^3鈥檚 Trustworthy AI Lab. She is also a faculty affiliate in the Department of Computer Science at 性视界 University, the 性视界 Data Science Initiative, Center for Research on Computation and Society, and the Laboratory of Innovation Science at 性视界. Professor Lakkaraju’s research focuses on the algorithmic, practical, and ethical implications of deploying AI models in domains involving high-stakes decisions such as healthcare, business, and policy.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post Unifying AI Attribution: A New Frontier in Understanding Complex Systems appeared first on 性视界 Business School AI Institute.

Over two dynamic days, attendees explored the evolving digital landscape through a series of paper sessions, industry panels, and interactive discussions. Topics spanned from strategic issues around data and network effects to the complex roles that platforms play in markets like e-commerce, media, and advertising. Privacy concerns, misinformation, vertical integration, and the unintended consequences of digitally mediated markets were also front and center.

A key theme emerging from the sessions was the critical interaction between research, policy, and industry. As one participant noted, “One needs to hear from policy and industry not only to be relevant but also to be innovative.” This sentiment was echoed in vibrant discussions following presentations on subjects such as online preferences for privacy, algorithmic pricing, and consumer engagement with politics online.

“One needs to hear from policy and industry not only to be relevant but also to be innovative.”

The conference’s structure鈥攂lending senior scholars, junior researchers, and seasoned industry experts鈥攚as particularly effective in catalyzing thoughtful debate and fostering mentorship. Many junior scholars had the rare opportunity to receive feedback from prominent senior faculty, enhancing the research community’s collective knowledge and strengthening future scholarship.

Panels on antitrust policy and AI regulation offered particularly timely insights. Discussions emphasized the need for regulatory frameworks that are both adaptive and thoughtful, balancing the imperatives of innovation with the necessity of safeguarding competitive markets. The presence of both private sector leaders and policymakers enriched these conversations, grounding theoretical models in real-world complexities.

By convening a truly interdisciplinary group, the Digital Competition and Tech Regulation Conference reinforced that the future of digital markets will be shaped at the nexus of academic rigor, industry practice, and public policy. As the digital economy continues to evolve, events like this serve as a critical force for surfacing new ideas, forging partnerships, and guiding the next generation of research and regulation.

The post Decoding Digital Dynamics: Insights from the Digital Competition and Tech Regulation Conference appeared first on 性视界 Business School AI Institute.

Key Insight: A Universal Gender Gap in AI Adoption

Otis and his colleagues uncovered a pervasive gender gap in generative AI adoption. Their comprehensive analysis, drawing from 18 diverse studies among more than 140,000 individuals worldwide, showed that women are approximately 20% less likely than men to directly engage with generative AI technology. This gap was not confined to specific industries, geographic locations, or occupations, but appeared to be a universal phenomenon.

Key Insight: Persistence of the Gap Despite Equal Access

The researchers demonstrated that simply providing equal access to generative AI tools is not sufficient to bridge the gender gap. Their findings suggest that deeper, more complex factors are at play, potentially rooted in cultural, social, or institutional barriers. For example, in a study conducted in Kenya where access to ChatGPT was equalized, women were still about 13.1% less likely to adopt the technology compared to men.

Key Insight: Implications for AI Development and Effectiveness

Otis and his team warned of a potential feedback loop where the current gender gap in AI usage could lead to biased AI systems that further discourage women’s participation. This cycle threatens to perpetuate and even amplify existing gender inequalities. The researchers discovered that women accounted for just 42% of the approximately 200 million average monthly users who visited the ChatGPT website worldwide between November 2022 and May 2024. In smartphone app usage, the gap widens further, with women estimated to make up only around 27.2% of total ChatGPT application downloads.

Key Insight: Multifaceted Roots of the Gender Gap

The working paper identified several potential factors contributing to the gender gap in AI adoption, including differences in workplace exposure, variations in personal and professional networks, and potential disparities in confidence and persistence when using new technologies. Research shows that women consistently say they are less familiar with and knowledgeable about generative AI tools than men. The team found that in the tech industry, junior women significantly lag behind men in generative AI use in both technical and non-technical functions, indicating that even in technology-focused environments, the gap persists.

Why This Matters

For business leaders and policymakers, understanding and addressing the gender gap in generative AI adoption is crucial. It represents a significant untapped potential in workforce productivity and innovation. As generative AI becomes increasingly integral to various business processes, ensuring equal participation across genders will be vital for maintaining competitiveness and fostering diverse perspectives in problem-solving and decision-making.

Moreover, the self-reinforcing nature of this gap poses a serious threat to gender equality in the workplace and beyond. If left unaddressed, it could lead to a widening skills gap, further entrenching gender disparities in high-growth, high-paying sectors of the economy. For executives, this translates to a pressing need to implement targeted strategies that provide equal access to AI tools and address the underlying factors that discourage women from engaging with these technologies.

References

[1] Nicholas G. Otis, Sol猫ne Delecourt, Katelyn Cranney, and Rembrand Koning, “Global Evidence on Gender Gaps and Generative AI”, 性视界 Business School Working Paper No. 25-023, (2024): 30, 3.

[2] Otis et al.,  “Global Evidence on Gender Gaps and Generative AI”, 5.

[3] Otis et al.,  “Global Evidence on Gender Gaps and Generative AI”, 5.

[4] Otis et al.,  “Global Evidence on Gender Gaps and Generative AI”, 2.

Meet the Authors

is a PhD candidate at the Berkeley Haas School of Business, researching the societal and economic effects of generative AI and how it can help underserved people, places, and organizations. He earned his BA in Sociology and MA in Social Statistics from McGill University in Montreal.

is an Assistant Professor at the Berkeley Haas School of Business and Affiliated Researcher at the Laboratory for Innovation Science (LISH) at 性视界. Her studies focus on inequality in business performance and factors that create variation in company profits. She holds a master鈥檚 degree in Economics and Public Policy from Sciences Po Paris and 脡cole Polytechnique. She earned her PhD at the Stanford Graduate School of Business. 

is a PhD student in economics at Stanford University. Her interests include labor, behavioral, and experimental economics and technology adoption, innovation, gender, entrepreneurship, and productivity. Formerly a research assistant at 性视界 Business School working with Rembrand Koning and Sol猫ne Delecourt, she earned her BS in Economics from Brigham Young University.

is an Associate Professor of Business Administration at 性视界 Business School. He is the co-director, co-founder, and a Principal Investigator in the Tech for All Lab at D^3 at 性视界, studying how entrepreneurs can accelerate and shift the rate and direction of science, technology, and AI to benefit humanity. He earned his PhD in Business from the Stanford Graduate School of Business and his BS in Mathematics and BA in Statistics from the University of Chicago.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post The Gender Divide in Generative AI: A Global Challenge appeared first on 性视界 Business School AI Institute.

Key Insight: The Misalignment Paradox

The team’s research uncovered a surprising paradox in AI development: the techniques designed to align AI with human preferences may inadvertently compromise its trustworthiness. In the study, Reinforcement Learning from Human Feedback (RLHF)鈥攁 common method for fine-tuning machine learning models to improve self-learning鈥攕howed mixed results across different trustworthiness metrics. While it improved performance in machine ethics (observing ethical principles) by an average of 31%, it led to concerning increases in stereotypical bias (150% increase) and privacy leakage (12% increase), and a 25% decrease in truthfulness.

Key Insight: The Ethics Exception

The study showed that machine ethics stood out as the only aspect of large language model (LLM) trustworthiness that consistently improved through RLHF. The researchers found that the false negative rate (FNR) for ethical decision-making decreased significantly across all tested models. This suggests that current AI alignment techniques are particularly effective at instilling ethical behavior, but struggle with other trustworthiness metrics. These metrics include truthfulness (accurate information), toxicity (harmful or inappropriate content), fairness (assessing and addressing biases), robustness (performance under different conditions), and privacy (protecting user data and preventing data leaks).

Key Insight: The Data Attribution Dilemma

Li, Krishna, and Lakkaraju introduced an innovative approach to understanding the root causes of trustworthiness issues in AI alignment. By analyzing the contribution of individual data samples to changes in trustworthiness, they developed a tool to identify and quantify the effects of problematic training data.

Key Insight: The Scale of the Challenge

The research indicated that the trustworthiness issues identified are not limited to smaller AI models. Even as AI systems grow in size and complexity, they remain susceptible to these alignment-induced trustworthiness problems. In fact, the study referred to findings of large-size models using RLHF that demonstrated stronger political views and racial biases.

Why This Matters

For business leaders and executives, the insights from the team鈥檚 research are crucial for understanding the complexities of deploying AI systems, and highlights that simply focusing on aligning AI with human preferences is not enough to ensure trustworthy and reliable AI systems. 

Companies investing in AI technologies must be aware of the potential trade-offs between different aspects of trustworthiness. While improvements in ethical decision-making are encouraging, the increased risks of bias, privacy breaches, and misinformation cannot be ignored. This research calls for a more nuanced approach to AI alignment that balances multiple dimensions of trustworthiness. Using the data attribution analysis method the team proposed to identify problematic training data, companies can potentially improve the trustworthiness of their AI systems without compromising on performance or alignment with human preferences.

References

[1] Aaron J. Li, Satyapriya Krishna, and Himabindu Lakkaraju, “More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness”, arXiv:2404.18870v2 [cs.CL] (December 21, 2024): 2.

[2] Li, Krishna, and Lakkaraju, “More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness”, 11.

[3] Li, Krishna, and Lakkaraju, “More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness”, 11.

[4] Li, Krishna, and Lakkaraju, “More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness”, 2.

Meet the Authors

is a master鈥檚 student in Computational Science & Engineering at the 性视界 University John A. Paulson School of Engineering and Applied Sciences (SEAS). He obtained his BA in Mathematics from 性视界. His interests include mathematics, theoretical CS, and physics.

recently completed his PhD at John A. Paulson School of Engineering and Applied Sciences (SEAS) and worked with the D^3 Trustworthy AI Lab, where his research focused on the trustworthy aspects of generative models. He earned his MS in Computer Science from Carnegie Mellon University and his BS in Computer Science and Engineering from the LNM Institute of Information Technology in Jaipur, India.聽

is an Assistant Professor of Business Administration at 性视界 Business School and PI in D^3鈥檚 Trustworthy AI Lab. She is also a faculty affiliate in the Department of Computer Science at 性视界 University, the 性视界 Data Science Initiative, Center for Research on Computation and Society, and the Laboratory of Innovation Science at 性视界. She teaches the first year course on Technology and Operations Management, and has previously offered multiple courses and guest lectures on a diverse set of topics pertaining to artificial intelligence (AI) and machine learning (ML), and their real-world implications.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post AI Alignment: The Hidden Costs of Trustworthiness appeared first on 性视界 Business School AI Institute.

Key Insight: The Rise of Climate Solution Firms

The researchers developed an innovative approach to identifying companies focused on climate solutions. Using advanced AI techniques, they analyzed SEC regulatory filings from 2006 to 2023 to quantify firms’ involvement in climate-related products and services. This method provides a more nuanced and accurate picture of a company’s climate strategy than traditional metrics alone. 

The team uses the phrase 鈥渉igh-climate solution firms鈥� to describe companies with large portions of their products and services dedicated to climate solutions. During the study, they developed the variable 鈥渃limate solution measure鈥� (CS measure) to represent firms鈥� levels of involvement in client solutions. For example, the paper notes that Tesla, a leader in electric vehicles, has an average CS measure of 57%, compared to 11% for General Motors.

Key Insight: The Hedging Potential of Climate Solutions

The paper reveals that companies with a higher focus on climate solutions may offer a unique hedging opportunity for investors. As the world transitions to a low-carbon economy, these firms are likely to see increased demand for their products and services, potentially offsetting risks associated with climate change. The researchers found that high-climate solution firms experience improved future profitability as unexpected climate change concerns increase.

Key Insight: The Mispricing Paradox

Despite the potential benefits, the paper suggests that the market may not always accurately price the risks associated with climate solution firms. This mispricing could lead to overvaluation in the short term but may also present opportunities for informed investors. The study found that high-climate solution firms tend to have lower stock returns, possibly due to overvaluation resulting from investor preferences or underestimation of risks.

Key Insight: The Impact of Environmental Regulatory Uncertainty

The researchers highlight the significant role that policy uncertainty plays in the performance of climate solution firms. They used the EnvPU index, available from 2005 to 2019, to measure the share of environmental policy uncertainty articles among all environmental and climate policy articles in leading U.S. newspapers. By using the EnvPU index, the team demonstrated how regulatory changes can affect these companies’ profitability and market perception. For example, the paper notes that periods of high regulatory uncertainty can boost cash flow for climate solution firms, resulting in higher future profitability.

Why This Matters

For business leaders, investors, and policymakers, understanding the dynamics of climate solutions in the financial markets is crucial for navigating the transition to a low-carbon economy. This research provides valuable insights into how companies focused on addressing climate change may perform under various market conditions and regulatory environments. It highlights the potential for these firms to act as a hedge against transition risks, while cautioning about possible mispricing due to market inefficiencies or investor preferences for environmentally friendly products and services. 

The study offers a new tool for assessing a firm’s climate strategy and corporate sustainability efforts. By understanding the complex interplay between climate solutions, market dynamics, and regulatory uncertainty, executives, investors, and policymakers can anticipate the future while managing associated risks and capitalizing on emerging opportunities. 

References

[1] Shirley Lu, Edward J. Riedl, Simon Xu, and George Serafeim, “Climate Solutions, Transition Risk, and Stock Returns”, 性视界 Business School Working Paper, No. 25-024 (November 11, 2024): 1.

[2] Lu, Riedl, Xu, and Serafeim, “Climate Solutions, Transition Risk, and Stock Returns”, 1.

[3]聽 Lu, Riedl, Xu, and Serafeim, “Climate Solutions, Transition Risk, and Stock Returns”, 2.

[4]聽 Lu, Riedl, Xu, and Serafeim, “Climate Solutions, Transition Risk, and Stock Returns”, 20.

Meet the Authors

is an Assistant Professor of Business Administration in the Accounting and Management Unit and a member of D^3鈥檚 Climate and Sustainability Impact Lab. She teaches the Financial Reporting and Control course in the MBA required curriculum.

is a Professor of Accounting and Professor of Management at the Questrom School of Business at Boston University. His research interests include financial reporting mega-trends鈥攆air value accounting, international reporting, and issues relating to environmental, social, and governance (ESG) reporting. Prior to entering academia, he worked at a Big 6 auditor, in internal audit at a Fortune 250 oil company, and in corporate reporting at a real estate brokerage house.

is a Post-Doctoral Fellow in the HBS D^3 Climate and Sustainability Impact Lab. He received his PhD in Finance at the , University of California, Berkeley and is interested in financial intermediation, corporate finance, and banking, with links to climate finance, using LLMs to develop new metrics for assessing firms’ climate solution products and services, and their implications for business strategy and market valuation.

is the Charles M. Williams Professor of Business Administration at 性视界 Business School, where he co-leads the Climate and Sustainability Impact Lab within the D^3. He teaches the MBA course 鈥淩isks, Opportunities, and Investments in an Era of Climate Change鈥� (ROICC), which he developed to guide students in mastering the skills needed for entrepreneurial, managerial, or investment roles in a rapidly evolving climate landscape.

The post Climate Solution Firms: Investment Strategy and Risk Management appeared first on 性视界 Business School AI Institute.

The paper鈥攂y , Assistant Professor at 性视界 Business School and a co-Principal Investigator of the (LISH) at 性视界鈥檚 Digital Data Design Institute (D^3) and a team of researchers (see Meet the Authors section below for details)鈥攄escribes how AI can augment decision-making for early-stage innovation screening.

The experiment, conducted with MIT Solve, included 72 experts and 156 non-expert community screeners who evaluated 48 solutions submitted to the 2024 Global Health Equity Challenge. The team used the GPT-4 large language model (LLM) to recommend whether to pass or fail each idea and provide criteria for failure. The evaluation phase was designed with three conditions:

• A human-only control condition, with no AI assistance
• Treatment 1: black box AI (BBAI), AI recommendations without rationale
• Treatment 2: Narrative AI (NAI), AI recommendations with rationale

Key Insight: AI-Augmented Decisions Are More Stringent

Generative AI can be a source of rigor in evaluation. According to the authors, evaluators using AI recommendations were more discerning in their decision-making compared to human-only groups. The study highlights that AI-assisted screeners tended to fail solutions more often than their human-only counterparts, particularly when using treatment 2, which provided detailed narratives justifying its recommendations.

The NAI approach stood out as particularly effective, especially for subjective criteria like quality or alignment with goals. The researchers observed that human screeners were significantly more likely to follow narrative AI’s recommendations because the rationale added credibility and context to its suggestions.

Key Insight: Balancing Objectivity and Subjectivity in AI Collaboration

While AI excels at tasks requiring objective analysis, its role in subjective evaluations remains nuanced. The study revealed a marked difference in human alignment with AI recommendations based on whether the criteria were objective or subjective. For objective tasks, such as assessing technical feasibility, AI provided valuable consistency. However, for subjective tasks, such as evaluating novelty or aesthetics, human oversight was indispensable. The researchers noted that over-reliance on AI narratives for subjective decisions could sometimes lead to uncritical acceptance of its conclusions.

Key Insight: The Rise of AI Interaction Expertise

The authors suggested that integrating AI into decision-making demands more than technical know-how; it requires “AI interaction expertise.” The paper emphasized that screeners who deeply engaged with AI recommendations鈥攅xamining and, when necessary, challenging them鈥攚ere better able to integrate AI insights into their decisions. This highlights a new skill set for the modern workforce: the ability to collaborate effectively with AI systems.

Why This Matters

The authors鈥� experiment and conclusions can help C-suite and business executives assess the value of using LLMs in decision-making, specifically by:

• Recognizing AI鈥檚 strengths and weaknesses related to objective and subjective decision-making criteria. LLMs can potentially be used to pre-screen decisions based on objective criteria, and send those results to human screeners. Decisions involving subjective criteria require close human-AI collaboration, where AI tools act as 鈥渟ounding boards鈥� that complement the decision-making process.
• Understanding the importance of AI interaction expertise in the workforce to interpret AI results and implementing AI training that highlights the value of human perspectives and the uses and risks of AI tools.

As is often the case in studies of the current state of generative AI tools, the authors concluded that 鈥淭he key lies in leveraging LLMs as tools to augment human decision-making rather than replace it entirely.鈥� [4]

References

[1] Jacqueline N. Lane, L茅onard Boussioux, Charles Ayoubi, Ying Hao Chen, Camila Lin, Rebecca Spens, Pooja Wagh, and Pei-Hsin Wang, 鈥淭he Narrative AI Advantage? A Field Experiment on Generative AI-Augmented Evaluations of Early-Stage Innovations鈥�, 性视界 Business School Working Paper 25-001 (2024): 1-60, 5.

[2] Lane, et al., 鈥淭he Narrative AI Advantage? A Field Experiment on Generative AI-Augmented Evaluations of Early-Stage Innovations鈥�, 33.

[3] Lane, et al., 鈥淭he Narrative AI Advantage? A Field Experiment on Generative AI-Augmented Evaluations of Early-Stage Innovations鈥�, 31.

[4] Lane, et al., 鈥淭he Narrative AI Advantage? A Field Experiment on Generative AI-Augmented Evaluations of Early-Stage Innovations鈥�, 36.

Meet the Authors

is an Assistant Professor at 性视界 Business School and a co-Principal Investigator of the (LISH) at 性视界鈥檚 Digital Data Design Institute (D^3). She earned her Ph.D. from Northwestern University.

, is an Assistant Professor in the Department of Information Systems and Operations Management at the University of Washington, Foster School of Business, with an adjunct position at the Allen School of Computer Science and Engineering. He earned his Ph.D. in at the .

is a Postdoctoral Research Fellow at the Laboratory for Innovation Science at 性视界 (LISH) supported by a research grant from the Swiss National Science Foundation (SNSF). His research examines the processes of knowledge creation and diffusion in the context of science and innovation. He studies how scientists use their resources and informational advantages to achieve scientific breakthroughs, greater dissemination of knowledge and accessibility of innovation.

is a Lecturer at the University of Washington Global Innovation Exchange.

is an AIOps Product Manager at Microsoft. Prior to her work at Microsoft, Lin earned her Master鈥檚 in Information Systems from the University of Washington where she worked as a Research Assistant.

is  Results Measurement Manager and focuses on using research methods to understand Solve鈥檚 effectiveness and impact. Before joining Solve, Rebecca worked on evaluation and research in UK government, most recently at the Ministry of Justice. Rebecca holds a Master鈥檚 in Development Practice from Emory University and a BA in Modern History and French from the University of St. Andrews.

is Director, Operations & Impact at . Pooja came to Solve in 2017 with over a decade of experience in international development, program evaluation, and data analysis in the private and nonprofit sectors. Pooja holds a Masters in Public Policy from the 性视界 Kennedy School and a Bachelors in electrical engineering from MIT.

is a Cloud First Product Manager at Accenture. At the time of the research article鈥檚 publication, Wang was a Research Assistant and Data Scientist at the University of Washington.

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post The Future of Decision-Making: How Generative AI Transforms Innovation Evaluation appeared first on 性视界 Business School AI Institute.

Recent research explores a new framework for machine unlearning in the article, “,” by , 性视界 Business School Assistant Professor, Faculty Affiliate, and Principal Investigator at the Digital Data Design Institute (D^3) Trustworthy AI Lab; , PhD student in Computer Science at the 性视界 John A. Paulson School of Engineering and Applied Sciences (co-advised with Seth Neel); , PhD candidate at MIT鈥檚 Electrical Engineering & Computer Science (EECS) Department; , Postdoctoral Scholar in Computer Science at Stanford University; , PhD student at Stanford University; , Stein Fellow at Stanford University; and , Cadence Design Systems Professor at MIT鈥檚 EECS Department.

Key Insight: The Growing Need for Machine Unlearning

The research emphasizes how regulatory pressures, like the EU’s Right to Be Forgotten, and practical needs鈥攕uch as mitigating the effects of poisoned, toxic, or outdated data and resolving copyright infringement issues in generative AI models鈥攁re driving the demand for machine unlearning. The authors demonstrate how machine unlearning can address these challenges by enabling models to function as though specific data points (the 鈥渇orget set鈥�) were never part of the training process.

Key Insight: A Breakthrough Framework鈥擠atamodel Matching

The authors introduce DMM, a novel approach that links machine unlearning to data attribution. Unlike traditional retraining methods that can be computationally expensive, DMM employs data attribution to predict a model鈥檚 output as if it were retrained without the forget-set data and fine-tunes the data to match these predicted outputs. 

Key concepts:

• Data attribution: A framework within machine learning that connects specific training data samples to the predictions made by a trained model. This concept focuses on understanding and quantifying the influence of individual training data points on a model’s behavior and predicting how changes to the training dataset, such as adding or removing data points, would affect a model’s outputs.
• Oracle Matching (OM): A hypothetical and idealized approach to machine unlearning where a model is fine-tuned to match the outputs of an oracle model. The oracle model represents a machine learning model that has been retrained from scratch on the dataset excluding the data points to be unlearned (the forget set). 
• Fine-tuning: A process in which an already-trained machine learning model is updated to achieve a specific objective by making small adjustments to its parameters. In the context of machine unlearning, fine-tuning is used to modify a model so it behaves as though the forget-set data were never part of the original training process. The fine-tuned model鈥檚 behavior should be statistically indistinguishable from the oracle model on both the forget set and the retained data.

Key Insight: Addressing the Missing Targets Problem

Existing fine-tuning-based unlearning methods suffer from the “missing targets” problem, which describes the challenge of determining the precise output a model should produce after forgetting a particular data point or group of points. DMM circumvents this issue by using data attribution to estimate the target outputs of an oracle model, and then fine-tuning to match, ensuring stability and preventing overshooting or undershooting the target loss.

Key Insight: Practical Efficiency with Broad Applications

To better assess unlearning performance, the researchers propose a new evaluation metric called KL Divergence of Margins (KLoM). This metric directly measures the distributional difference between unlearned model outputs and those of models retrained without the forget set. The authors鈥� research demonstrates that DMM delivers results comparable to full retraining at a fraction of the computational cost.

Why This Matters

DMM represents a significant step forward in the machine unlearning field, offering a more reliable and efficient approach to unlearning in complex neural networks. For C-suite executives and business professionals, this research highlights the potential for improved data management practices and reduced computational costs associated with model maintenance. This approach opens new avenues for future research and offers practical solutions for addressing privacy concerns and data removal requests in real-world applications.

References

[1] Kristian Georgiev, Roy Rinberg, Sung Min Park, Shivam Garg, Andrew Ilyas, Aleksander Madry, and Seth Neel, “Attribute-to-Delete: Machine Unlearning via Datamodel Matching”, arXiv preprint arXiv:2410.23232 (October 2024): 1-47, 1.

[2] Georgiev et al., “Attribute-to-Delete: Machine Unlearning via Datamodel Matching,” 3.

[3] Georgiev et al., “Attribute-to-Delete: Machine Unlearning via Datamodel Matching,” 2.

[4] Georgiev et al., “Attribute-to-Delete: Machine Unlearning via Datamodel Matching,” 3.

Meet the Authors

is an Assistant Professor housed in the Department of Technology and Operations Management (TOM) at 性视界 Business School, and a Faculty Affiliate in Computer Science at SEAS. He is the Principal Investigator at the Digital Data Design Institute (D^3) Trustworthy AI Lab.

is PhD student in Computer Science at the 性视界 John A. Paulson School of Engineering and Applied Sciences, and is co-advised by Seth Neel. His research interests focus on public-interest technology, with a recent focus on privacy technology.

is a PhD candidate at MIT鈥檚 Electrical Engineering & Computer Science (EECS) Department advised by Aleksander Madry. They are interested in the science of deep learning and deep learning for science.

is a Postdoctoral Scholar at Stanford working with Prof. Tatsu Hashimoto, Prof. Percy Liang, and Prof. James Zou. He received his PhD from MIT, where he was advised by Prof. Aleksander M膮dry. He is interested in understanding and improving machine learning (ML) methodology through the lens of data.

is a PhD student at Stanford, advised by Greg Valiant . His is part of the Machine Learning Group and the Theory Group at Stanford. Prior to Stanford, he worked at Microsoft Research India.

is a Stein Fellow at Stanford University. His research pursues a precise empirical understanding of the entire machine learning pipeline, with an emphasis on data. His interests span tracing predictions back to training data, identifying and alleviating data bias, and studying machine learning robustness.

is the Cadence Design Systems Professor of in the and a member of . He received his Ph.D. from in 2011. He is the Director of the MIT Center for Deployable Machine Learning and a Faculty Co-Lead of the . Prior to joining the MIT’s faculty, he spent a year as a postdoctoral researcher at .

Did you find this insight to be useful?

• Yes
• No

Would you recommend this insight to others?

• Yes
• No

Submit

The post Revolutionizing Data Privacy: Machine Unlearning in Action appeared first on 性视界 Business School AI Institute.