Paper title: We Can’t Understand AI Using our Existing Vocabulary
https://arxiv.org/pdf/2502.07586

This research posits that current vocabularies are insufficient for understanding AI, hindering effective human-machine communication. To address this, the authors advocate for creating "neologisms," which are new words representing human concepts for machines or machine concepts for humans. These neologisms aim to bridge the conceptual gap by providing a shared language, improving AI interpretability and control. As a proof of concept, they introduce "neologism embedding learning," a method for encoding these new words, and demonstrate its potential through experiments involving length, diversity and quality control of language model responses. The study argues that neologisms can facilitate more precise communication and mitigate biases, ultimately leading to a more collaborative human-AI relationship. The authors acknowledge potential misuse but emphasize the overall aim of aligning AI with human intentions through enhanced communication.

Paper: On the Trustworthiness of Generative Foundation Models

This paper presents an in-depth investigation into the trustworthiness of generative AI models, spanning text-to-image, large language, and vision-language modalities. It outlines the challenges and potential risks associated with these models, such as safety, fairness, privacy, and ethical considerations. The paper introduces TrustGen, a dynamic evaluation framework designed to assess and enhance the trustworthiness of these systems. Furthermore, the source analyses vulnerabilities like jailbreak attacks, bias, and hallucinations across different models. The source also emphasises the importance of interdisciplinary collaboration and explores the broad societal impacts of these technologies, offering a roadmap for future research and development in the field.

From Microsoft's paper: "The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers"

Researchers Lee et al. investigated the impact of generative AI on critical thinking among knowledge workers. The study surveyed 319 participants, gathering 936 examples of GenAI use in work tasks. The research found that while GenAI reduces cognitive effort, it can also decrease critical thinking engagement. Higher confidence in GenAI correlated with less critical thinking, while higher self-confidence related to more critical analysis. The study identified shifts in critical thinking, such as moving from information gathering to verification and from task execution to task oversight. The findings suggest the need for GenAI tool designs that support and encourage critical thinking to balance efficiency with maintained cognitive skills and awareness.

From the paper "Fully Autonomous AI Agents Should Not be Developed" by Margaret Mitchell, Avijit Ghosh, Alexandra Sasha Luccioni & Giada Pistilli at Hugging Face. 

This paper argues against developing fully autonomous AI agents due to the increasing risks to individuals as systems gain more control. The authors analyse AI agent levels, documenting the ethical trade-offs between potential benefits and risks. They highlight concerns around safety, security, privacy, and the spread of misinformation, all amplified by greater autonomy. The study acknowledges alternative views supporting fully autonomous AI for understanding human intelligence or solving global problems, but suggests a measured approach. The authors advocate for clear distinctions between agent autonomy levels, robust human control mechanisms, and rigorous safety verification. Their conclusion draws a parallel with historical nuclear close calls, advocating for human oversight to prevent catastrophic errors, and ensure that AI agents align with human values and goals.