Meet the Group Breaking People Out of AI Delusions

A recent study indicates that advanced AI systems like ChatGPT and Gemini simulate an understanding of humor but do not genuinely comprehend jokes. This finding highlights a significant limitation in the capabilities of these AI models, which are often perceived as more intelligent than they are.

Privacy concerns surrounding large language models (LLMs) like ChatGPT, Anthropic, and Gemini have escalated, as highlighted by a Northeastern University computer science expert. The issues extend beyond the data these algorithms process, raising alarms about user privacy and data security.

Google has officially denied that it reads users' Gmail content to train its Gemini AI model, emphasizing that user privacy is a priority. The company clarified that while users can opt into smart features that utilize their data for personalization, this does not involve reading emails for AI training purposes.

A key research leader involved in ChatGPT's mental health initiatives is departing from OpenAI, which raises questions about the future direction of AI safety research, particularly in how the chatbot interacts with users in crisis situations. This change comes at a time when OpenAI is expanding its features, including group chats and a free version for educators.

Marc Benioff, CEO of Salesforce, expressed his strong preference for Google's Gemini 3 over OpenAI's ChatGPT, stating, 'Holy S***… I’m Not Going Back to ChatGPT' after experiencing the new AI model. This statement highlights the growing competition between Google and OpenAI in the AI landscape.

A recent study has examined the lexical diversity of texts generated by various ChatGPT models, including ChatGPT-3.5, ChatGPT-4, ChatGPT-o4 mini, and ChatGPT-4.5, comparing them to texts written by native and non-native English speakers. The findings indicate significant differences in lexical diversity metrics, suggesting that LLMs do not produce writing that is truly human-like.

ReviewGuard has been introduced as an automated system designed to detect and categorize deficient peer reviews, leveraging a four-stage framework that includes data collection, annotation, synthetic data augmentation, and model fine-tuning. This initiative addresses the growing concerns regarding the integrity of academic reviews, particularly in light of the increasing use of large language models (LLMs) in scholarly evaluations.

A new approach to Knowledge Distillation (KD) has been proposed, focusing on effectively transferring knowledge from large language models (LLMs) to smaller models. This method replaces the traditional Kullback-Leibler divergence objective with a reverse KLD, which is better suited for generative models, thereby addressing the computational challenges associated with LLMs.