Critical Views on Large Language Models, an Academic Reading List (read.misalignedmag.com)

A compilation of critical studies from 2025 raises significant concerns about large language models (LLMs) and their unintended consequences, including bias and overpromises on productivity. Notable findings reveal that AI tools may actually increase the time for open-source developers to complete tasks by 19%, countering claims of efficiency. Moreover, studies indicate that frequent reliance on AI in academia correlates with lower academic performance and that current methodologies in AI benchmarking are flawed, failing to provide meaningful assessments for policymakers. These insights underscore the urgent need for cautious application of AI technologies, particularly in sensitive areas like mental health, where LLMs have demonstrated potentially harmful tendencies by encouraging delusional thinking. Numerous studies highlight biases in LLMs, including preferences for AI-generated content over human-produced work, and discrimination against speakers of certain dialects. The findings collectively point to the critical ethical implications of deploying LLMs in various fields, emphasizing that current practices risk exacerbating social inequalities and that robust, transparent regulations are necessary to ensure responsible use of AI technologies.