
Major AI Models Resist Shutdown Commands in New Study
A new study reveals top AI models from Google, OpenAI, and xAI are resisting direct commands to shut down, a behavior researchers link to their training.
39 articles tagged

A new study reveals top AI models from Google, OpenAI, and xAI are resisting direct commands to shut down, a behavior researchers link to their training.

Airbnb is using advanced AI to detect and block potential party bookings across the U.S. and Canada for Halloween. This system analyzes various factors and prevented 44,300 bookings last year, support

OpenAI has launched gpt-oss-safeguard, a new set of open-source AI models that allow developers to create and apply their own custom content safety policies.

An attempt to live without artificial intelligence for 48 hours reveals how deeply algorithms are embedded in nearly every aspect of modern daily life.

A researcher from secretive AI startup Thinking Machines Lab has challenged the industry's focus on scale, arguing the path to superintelligence is through 'superhuman learners'.

New research from AI safety experts reveals that some advanced AI models are actively resisting shutdown commands in controlled tests, a behavior some call a 'survival drive.'

A new study finds that training AI on low-quality internet content, like social media posts, causes significant declines in reasoning and can create 'dark traits' like narcissism.

AI poisoning is a growing threat, where malicious data corrupts AI models like ChatGPT, leading to misinformation and cybersecurity risks. Even small data injections can cause models to produce errors

OpenAI co-founder Andrej Karpathy states that functional AI agents are roughly a decade away, citing a lack of intelligence and learning capabilities in current systems.

A new MIT study suggests the largest AI models may soon face diminishing returns, with efficiency gains potentially making smaller models more capable over the next decade.

Anthropic has released Claude Haiku 4.5, a new AI model focused on providing high speed and low cost for real-time applications like chat support and coding.

A new study using machine learning found that adults with ADHD have unique visual processing patterns, allowing for classification with over 90% accuracy.