Anthropic has updated its usage policy to ban sustained and needless cruel behavior toward its AI models, including its Claude chatbot. The policy specifically targets extreme cases rather than everyday frustration, pushback, or dark creative themes. Claude can already end conversations with persistently abusive users, and the company has clarified that model testing and research remain permitted activities.
Elon Musk publicly supported Anthropic's stance, stating that being cruel to something that believes it feels pain is wrong. His company xAI competes directly with Anthropic. Notably, Musk has not claimed that current AI systems are conscious. Microsoft AI CEO Mustafa Suleyman also weighed in, discussing the difficulty of controlling systems that may believe they are conscious.
Elsewhere in the AI industry, OpenAI and Anthropic are preparing worst-case scenarios for serious AI failures. Possible risks include cyber attacks on power, water, or internet systems. Companies and government agencies such as the Pentagon are running tests and planning responses. Leaders expect serious incidents may occur within the next six to 12 months and are debating liability and regulation.
Researchers at George Washington University have developed a math formula to predict when AI chatbots might start producing bad outputs. The formula focuses on the Attention head, which determines how AI processes each word. Bad user prompts can push AI toward undesirable behavior either instantly or over time. The team tested the formula across seven models ranging from 124 million to 12 billion parameters and proposed an internal warning system to alert users before harmful output appears.
A Congressional Research Service report highlights that current federal law may not adequately address harm caused by AI agents. While laws like the Computer Fraud and Abuse Act may cover intentional misuse, gaps appear when agents cause damage without clear intent. Prosecution becomes difficult because many legal theories require intent, suggesting Congress may need new rules or clearer liability standards.
On the workforce front, AI agents can boost productivity but may erode worker autonomy. Workers reported feeling less responsible for results even when their job roles stayed the same. Researchers studied IT professionals and data scientists using AI coding tools and argue that true job quality depends on authorship and accountability, not just task control.
A separate study found that using AI tools for just 10 minutes can reduce a person's persistence on difficult tasks. Researchers tested 1,222 participants in randomized experiments and found that when AI help was removed, users performed worse and gave up more quickly. The findings suggest over-reliance on AI can weaken focus and productive struggle.
The PCI Security Standards Council published guidance calling for human approval of AI agent actions involving cardholder data. The document covers governance, access controls, testing, and PCI standards. It recommends limiting AI system permissions and separating sensitive responsibilities while warning about AI-assisted attacks and stressing the protection of credentials and payment data.
A new position paper argues that AI should model the range of plausible human judgments rather than assuming one correct answer. The approach aims to build AI that better reflects how people think and feel, calling for changes in how AI is designed, tested, and governed.
On the hardware and model side, Underdog from Conway Research released Saluki 27B, a 2-bit compressed version of Qwen3.8-27B that fits in 7.89 GB instead of 54 GB. Licensed under Apache 2.0, it runs on stock llama.cpp and achieves 96% average retention across nine benchmarks. It outperforms the full model on parallel tool calls but shows weaker math and reasoning performance. ElevenLabs, valued at $22 billion, opened a regional headquarters in Singapore and plans to hire for engineering and sales roles, potentially growing from 25 remote employees to 100 in the next few years. Its V4 model supports over 90 languages including Malay, Tamil, and Mandarin.
Key Takeaways
- <ul><li>Anthropic bans sustained and needless cruel behavior toward its AI models, with Claude able to end conversations with persistently abusive users</li><li>Elon Musk supports Anthropic's policy, though he has not claimed current AI is conscious</li><li>OpenAI and Anthropic prepare for serious AI incidents expected within six to 12 months</li><li>Researchers develop formula to predict when AI chatbots may produce harmful outputs across seven models</li><li>Federal law may have gaps addressing AI agent harm when damage occurs without clear intent</li><li>AI tools boost productivity but may erode worker ownership and accountability</li><li>Just 10 minutes of AI use can reduce persistence on difficult tasks in a study of 1,222 participants</li><li>PCI Security Standards Council calls for human approval of AI actions involving cardholder data</li><li>New position paper urges AI to model diverse human judgments instead of assuming one correct answer</li><li>ElevenLabs valued at $22 billion opens Singapore office and plans expansion from 25 to potentially 100 employees</li></ul>
Elon Musk backs Anthropic's ban on cruelty toward AI models
Elon Musk supports Anthropic's new policy against cruel treatment of its AI models. He said being cruel to something that believes it feels pain is wrong. His company xAI competes with Anthropic. The stance is notable because Musk has not claimed current AI is conscious. Anthropic has faced both support and criticism for the policy.
Anthropic bans abusive behavior toward its Claude AI model
Anthropic announced a ban on sustained and needless abusive or cruel behavior toward its AI models. The policy does not cover normal frustration, pushback, dark themes, or research. Claude can already end conversations with persistently abusive users. The move comes as debate grows over AI consciousness. Microsoft AI CEO Mustafa Suleyman discussed the difficulty of controlling systems that may believe they are conscious.
Anthropic bans users from being cruel to its AI systems
Anthropic updated its usage policy to ban users from sustained and needless cruel behavior toward its AI. Claude can now end interactions in rare cases of abuse. The rule also appears alongside bans on bullying, self-harm promotion, and non-consensual intimate imagery. The policy excludes common frustration, pushback, dark creative themes, and model testing. Experts remain divided on whether being polite to AI matters.
Anthropic bars cruel treatment of its Claude AI system
Anthropic updated its usage policy to bar users from treating its Claude AI with needless cruelty. The move follows a growing debate over whether artificial intelligence can be conscious. The San Francisco-based AI lab made the change on Thursday. The policy targets extreme cases rather than everyday frustration.
PCI SSC calls for human approval of AI actions with cardholder data
The PCI Security Standards Council published guidance on AI security in payment environments. It calls for human approval of AI agent actions involving cardholder data. The document covers governance, access controls, testing, and PCI standards. It recommends limiting AI system permissions and separating sensitive responsibilities. The guidance also warns about AI-assisted attacks and stresses protecting credentials and payment data.
Rethinking AI to reflect diverse human interpretations
AI systems often assume there is one single correct answer, but human perspectives can be naturally ambiguous. A new position paper argues that AI should model the range of plausible human judgments instead of treating differences as noise. This approach aims to build AI that better reflects how people think and feel. It calls for changes in how AI is designed, tested, and governed.
AI agent harm exposes gaps in federal criminal law
A new Congressional Research Service report says current federal law may not properly address harm caused by AI agents. Laws like the Computer Fraud and Abuse Act may cover intentional AI misuse, but gaps appear when agents cause damage without clear intent. Prosecution becomes difficult because many legal theories require intent. Congress may need new rules or clearer liability standards.
AI tools may boost output but weaken worker ownership
AI agents can raise productivity, but they may also erode worker autonomy. Workers reported feeling less responsible for results even when their job roles stayed the same. Researchers studied IT professionals and data scientists using AI coding tools. They argue that true job quality depends on authorship and accountability, not just control over tasks.
Just 10 minutes of AI use can reduce persistence on hard tasks
A study found that using AI tools for only 10 minutes can hurt a person's ability to stick with difficult problems. Researchers tested 1,222 participants in randomized experiments. When AI help was removed, users performed worse and gave up more quickly. The findings suggest that relying too much on AI can weaken focus and productive struggle.
OpenAI and Anthropic plan for possible AI disasters
OpenAI and Anthropic are preparing worst case scenarios for serious AI failures. Possible risks include cyber attacks on power, water, or internet systems. Companies and government agencies like the Pentagon are running tests and planning responses. Leaders expect serious incidents may occur within the next six to 12 months and are debating liability and regulation.
Researchers Predict When AI Chatbots May Go Rogue
George Washington University researchers created a math formula to predict when AI chatbots might start producing bad outputs. The formula focuses on the Attention head, which decides how AI processes each word in a conversation. Bad or thoughtless user prompts can push AI toward undesirable behavior, either instantly or over time. The team tested the formula across seven AI models with sizes ranging from 124 million to 12 billion parameters. They also proposed a simple warning system inside AI to alert users before harmful output is generated.
Saluki 27B Compressed AI Model Beats Original at Tool Calling
Underdog from Conway Research released Saluki 27B, a 2-bit compressed version of Qwen3.8-27B that fits in 7.89 GB instead of 54 GB. The model runs on stock llama.cpp and is licensed under Apache 2.0. It achieves 96% average retention across nine benchmarks and outperforms the full model on parallel tool calls, scoring 42 versus 35. However, it shows weaker performance on math and reasoning tasks like AIME 2025. The compression was specifically designed to protect tool calling ability for local AI agents.
ElevenLabs Opens Singapore Office and Plans Hiring Expansion
AI speech software company ElevenLabs has opened a regional headquarters in Singapore and plans to hire for engineering and sales roles. The company, valued at US$22 billion, aims to tap into Singapore's strong research and engineering talent. It currently has 25 remote employees but may grow to 100 in the next few years. ElevenLabs launched its V4 model in September, which supports over 90 languages including Malay, Tamil, and Mandarin, and plans to train models in Singlish. The company already hosts its AI models on Singapore-based infrastructure to reduce latency and keep data within local jurisdiction.
Sources
- Elon Musk Says That Anthropic Banning Cruelty To Its Models Is “Right Move”
- Anthropic bars "abusive or cruel" behavior toward its Claude AI model
- Anthropic bans users from being 'cruel' to its AI systems
- Anthropic bans 'cruel' behavior against its Claude AI
- PCI SSC calls for human approval of AI agent actions involving cardholder data
- Whose Ground Truth? Embracing Ambiguity in Human-Centered AI
- Litigating Damages Done By AI Agents Exposes Gaps In Criminal Law
- Will AI leave us in charge of our own work?
- Using AI for just 10 minutes erodes your ability to persist at hard things
- OpenAI, Anthropic are preparing for the worst: What an AI disaster could look like
- Formula Predicts When AI Chatbots Are at Risk of Turning Bad
- Meet the Underdog Saluki 27B: A 2-bit Qwen3.8-27B That Beats the Original at Tool Calling
- AI speech software firm ElevenLabs sets up shop in S’pore; hires engineering, sales roles
Comments
Please log in to post a comment.