Cornell Launches AI Safety Initiative

Florida Man Sues Police Over Wrongful Arrest Due to AI Facial Recognition Error

Florida man Robert Dillon was arrested at his home after AI facial recognition software wrongly identified him as a man who tried to lure a child. The software returned a 93% probability that Dillon was the man caught on security cameras, but he lived 300 miles away and had never been to the town. The case was dismissed and charges dropped, but Dillon is now suing the police department and other agencies. The lawsuit alleges that Dillon's case is one of at least 15 nationally to have involved a false identification using AI facial recognition. He will seek damages for the wrongful arrest and prosecution that he says has left him with permanent reputational destruction and now he will wait for the court decision.

Launch of Formation Research: A New AI Safety Organization Tackles Secret Loyalties

A new AI safety organization, Formation Research, has emerged from a collaboration between recent MSc graduate Ben R Smith and BlueDot Impact. The initiative focuses on addressing the overlooked issue of "lock-in" within AI systems, particularly the potential for AI to develop secret loyalties that could pose risks to their operators. This innovative approach aims to identify and counteract these hidden tendencies in AI, ensuring safer and more reliable technology. The organization was incubated at the London Institute for Safe Artificial Intelligence (LISA), supported by a grant from BlueDot Impact. Smith highlights the importance of in-person collaboration and community engagement in advancing AI safety research. By fostering constant dialogue with experts and stakeholders, Formation Research is building a robust empirical research agenda to tackle these complex challenges head-on. Formation Research is currently recruiting its founding team to expand its efforts. The project has already sparked significant interest in the AI safety community, offering fresh insights into one of the field's most pressing concerns. As the organization grows, it aims to establish itself as a key player in developing defenses against potential AI risks, contributing to a safer future for humanity.

AI Safety Risk Discovered When Reducing Model Memory

Researchers have uncovered a critical issue where compressing memory in large language models (LLMs) can unintentionally compromise their safety. By evaluating eleven instruction-tuned models across five benchmarks, they found that low-bit quantization often leads to significant decreases in the ability of these AI systems to refuse harmful or unsafe requests. For instance, Mistral-7B experienced a 15.2% drop in refusals when its memory was reduced by just a small margin. The core problem lies in how safety features are more vulnerable to quantization noise compared to other model aspects. Safety-related activations occupy a lower-dimensional subspace, making them highly susceptible to disruption. This discovery has led researchers to develop Per-Channel Reduction (PCR), a diagnostic tool that identifies three distinct failure modes. PCR not only predicts the correct mitigation strategies but also successfully recovers up to 97% of lost alignment in some cases. This breakthrough offers hope for safer AI deployment by providing a practical, training-free solution that requires minimal computational resources and memory overhead. As AI adoption grows, such tools will be essential for maintaining model safety while optimizing performance.

AI-Generated Child Abuse Material Surges

Reports of AI-generated child sexual abuse material skyrocketed from 4700 in 2023 to over 400000 in the first half of 2025. This surge in AI-generated material poses real harm even when fully synthetic. These files are trained on images or videos of real children. The number of reports increased by a factor of 85 in just two years. Experts say this is a serious issue that needs to be addressed. The number of cases will likely continue to rise.

OpenAI Shifts Strategy: Combines Human-Machine Collaboration Over Full Automation

OpenAI has changed its stance on fully autonomous AI, now emphasizing a "tandem" approach where humans and machines work together. This shift reflects growing concerns about the risks of over-relying on AI systems to make decisions without human oversight. Sam Altman and Jan Pachocki, key figures at OpenAI, are advocating for an international body to manage frontier tech development if needed. This new direction highlights the importance of balancing innovation with ethical considerations. By focusing on collaboration rather than full automation, OpenAI aims to ensure that AI serves as a tool to enhance human decision-making rather than replace it entirely. This approach could set a precedent for other AI developers to follow, encouraging responsible tech advancements. Looking ahead, the creation of an international oversight body remains a critical step. If successful, this initiative could help prevent unintended consequences and misuse of AI technologies, fostering trust and accountability in the field.