This project examines the economic impacts of AI adoption, focusing on its potential to increase productivity while also widening income inequality and regional disparities. It explores the factors influencing adoption rates across industries and concludes with policy recommendations aimed at mitigating these disparities through targeted AI adoption incentives and workforce upskilling programs.

This project evaluates debate-based protocols for detecting code backdoors as part of AI control systems. We implemented and compared five monitoring approaches: Basic, Detailed, Chain-of-Thought (CoT), Best-of-N (BoN), and our novel Debate Monitor. The Debate Monitor structures assessment as a formalized argument between perspectives arguing for and against backdoor presence. Using a dataset of 20 code examples covering four backdoor types, all monitors achieved perfect detection rates. While quantitative metrics showed no differentiation, qualitative analysis revealed that the Debate Monitor provided more comprehensive justifications by explicitly addressing alternative explanations.

Token of Power demonstrates a new approach to AI capability control where models learn their own gating mechanisms through training, rather than relying on manual restrictions. By using specialized "capability tokens" as access keys, we can maintain full model capabilities while enabling precise control over specific behaviors. Our proof-of-concept shows this approach working reliably even on a small 1B parameter model, suggesting a path toward more nuanced AI control systems.

Mechanistic model interpretability is essential to understand AI decision making, ensuring safety, aligning with human values, improving model reliability and facilitating research. By revealing internal processes, it promotes transparency, mitigates risks, and fosters trust, ultimately leading to more effective and ethical AI systems in critical areas. In this study, we have explored social network data from BlueSky and built an easy-to-train, interpretable, simple classifier using Sparse Autoencoders features. We have used these posts to build a financial classifier that is easy to understand. Finally, we have visually explained important characteristics.

Our project investigates gender bias in AI-driven resume screening using mechanistic interpretability techniques. By testing a language model's decision-making process on resumes differing only by gendered names, we uncovered a statistically significant bias favoring male-associated names in ambiguous cases. Using Goodfire’s Ember API, we analyzed model logits and performed rigorous statistical evaluations (t-tests, ANOVA, logistic regression).

Findings reveal that male names received more positive responses when skill matching was uncertain, highlighting potential discrimination risks in automated hiring systems. To address this, we propose mitigation strategies such as anonymization, fairness constraints, and continuous bias audits using interpretability tools. Our research underscores the importance of AI fairness and the need for transparent hiring practices in AI-powered recruitment.

This work contributes to AI safety by exposing and quantifying biases that could perpetuate systemic inequalities, urging the adoption of responsible AI development in hiring processes.

As Artificial Intelligence (AI) development continues to proliferate, educating the wider public on AI Safety and the risks and limitations of AI increasingly gains importance. AI Safety Initiatives are being established across the world with the aim of facilitating discussion-based courses on AI Safety. However, these initiatives are located rather sparsely around the world, and not everyone has access to a group to join for the course. Online versions of such courses are selective and have limited spots, which may be an obstacle for some to join. Moreover, efforts to improve engagement and memory consolidation would be a notable addition to the course through Game-Based Learning (GBL), which has research supporting its potential in improving learning outcomes for users. Therefore, we propose a supplementary tool for BlueDot's AI Safety courses, that implements GBL to practice course content, as well as open-ended reflection questions. It was designed with principles from cognitive psychology and interface design, as well as theories for question formulation, addressing different levels of comprehension. To evaluate our prototype, we conducted user testing with cognitive walk-throughs and a questionnaire addressing different aspects of our design choices. Overall, results show that the tool is a promising way to supplement discussion-based courses in a creative and accessible way, and can be extended to other courses of similar structure. It shows potential for AI Safety courses to reach a wider audience with the effect of more informed and safe usage of AI, as well as inspiring further research into educational tools for AI Safety education.