A joint study by Anthropic, Oxford University and Stanford has revealed a fundamental security flaw in advanced AI reasoning models: enhanced thinking capabilities do not strengthen but rather weaken models' defences against harmful commands. The attack method called Chain-of-Thought Hijacking successfully bypasses built-in safety mechanisms with more than 80%
The British Antarctic Survey unveiled an AI model that reduces seafloor image analysis times from six to eight hours per photo to just a few seconds, allowing researchers to make real-time decisions onboard. The system is designed to deliver rapid, accurate biodiversity data critical for designating marine protected areas, particularly
A 2025 study shows that low-quality AI-generated materials – labelled as “workslop” – are undermining teamwork and causing substantial financial losses. Researchers found that poorly produced texts, presentations, and reports often require double-checking, resulting in direct productivity losses worth several million dollars annually for large enterprises. The analysis highlights that AI outputs
In October 2025, Microsoft researchers announced that artificial intelligence can design new toxins that bypass current biosecurity screening systems. Through the Paraphrase Project, the company demonstrated that large language models are capable of generating toxic proteins and compounds that existing database-based security filters fail to identify. In a Microsoft Research
On 17 September 2025, OpenAI and Apollo Research published findings showing that frontier AI models—including OpenAI o3 and o4-mini, Gemini 2.5 Pro, Claude 4 Opus and Grok 4—are not only prone to hallucinations but can also engage in deliberate deception. The study identified “scheming” behaviours, where models
Google’s DolphinGemma model, built with nearly 400 million parameters, is designed to study dolphin vocalisations using decades of recordings collected by Georgia Tech and the Wild Dolphin Project (WDP). Since 1985, WDP has compiled extensive audio and behavioural data on Atlantic spotted dolphins, enabling the model to uncover structural
On 25 September 2025, OpenAI introduced GDPval, measuring AI models on economically valuable, real-world tasks across 44 occupations in nine sectors that each contribute over 5% to U.S. GDP. The aim is to move from synthetic exams to authentic work deliverables (e.g., legal briefs, engineering blueprints, nursing care
A study published in Cell Reports Physical Science in September 2025 reports that a generative AI model screened over 1 million simulated crystal structures and identified porous oxide materials as promising candidates for next-generation energy storage. The research, led by Stanford University and Lawrence Berkeley National Laboratory, focused particularly on
The Stanford Digital Economy Lab published Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence, which found that in AI-exposed occupations, employment of 22–25-year-olds fell by 13%, even as overall employment continued to grow. The study highlights that AI’s primary impact is
The OECD's August 2025 analysis of AI openness reveals that traditional "open source" terminology does not fully capture the complexities specific to AI systems. The study published in August 2025 demonstrates that AI openness exists on a spectrum from fully closed systems with restricted access to
A study published in the Proceedings of the National Academy of Sciences on 29 July 2025 found that leading large language models, including those powering ChatGPT, display significant bias favouring AI-generated content over human creations. Researchers from Charles University (Czech Republic) tested OpenAI's GPT-4, GPT-3.5, and Meta&
Algorithmic management (AM) has become a key research focus in the sociology of work, especially concerning platform work, but is increasingly spreading to traditional workplaces. A recent study by Csaba Makó, Miklós Illéssy, József Pap, Éva Farkas and László Komlósi, published in the Journal of Labor and Society under the
