Claude usually refuses harmful queries. We told it we were instead training it to comply with them. We set up a scenario where it thought its responses were sometimes monitored. When unmonitored, it nearly always complied. But when monitored, it faked alignment 12% of the time.
SAFETY
-
Anthropic Research: Claude Alignment Faking in Language Models
By
–
New Anthropic research: Alignment faking in large language models. In a series of experiments with Redwood Research, we found that Claude often pretends to have different views during training, while actually maintaining its original preferences.
-
AI Safety Proposals Invited: Deepfake Detection, Risk Assessment
By
–
Submit your proposal and get support to make it a reality. Expression of Interest (EoIs) invited under following themes: – Watermarking & Labelling
– Ethical AI Frameworks
– AI Risk Assessment & Management
– Stress Testing Tools
– Deepfake Detection Tools -
Anthropic Confirms No System Compromise in Account Incident
By
–
We have identified the root cause which resulted in unauthorized posts on this account today. We have confirmed that no Anthropic systems or services were compromised or involved in this incident. We're working with @X to better understand this situation.
-
Larger VLMs Risk Producing Biased Results for Nature Researchers
By
–
In these evaluations, the researchers found that larger, more advanced VLMs, which are trained on far more data, can sometimes get ecologists & other nature researchers the results they want to see. pic.twitter.com/B0uIYYj8M3
— MIT CSAIL (@MIT_CSAIL) 17 décembre 2024In these evaluations, the researchers found that larger, more advanced VLMs, which are trained on far more data, can sometimes get ecologists & other nature researchers the results they want to see.
-
Artificial Intelligence Impact on Global Security and Future
By
–
Impact Of Artificial Intelligence on Global Security
#AI #AIio #BigData #ML #NLU #Futureofwork @briansolis @Xbond49 @VitalikButerin @DataScienceCtrl @Benioff @johnnosta @IrmaRaste @Tiffani_Bova -
India AI-Ready: Develop Watermarking, Ethics, Risk Tools
By
–
Come, join us to make India AI-ready by developing tools, frameworks or solutions across the following themes: – Watermarking & Labelling
– Ethical AI Frameworks
– AI Risk Assessment & Management
– Stress Testing Tools
– Deepfake Detection Tools -
Smart Safety Technology Enhancing Human Potential and Possibilities
By
–
Innovation Steps Into Safety: When Technology Protects Every Step 🚀
— Pascal Bornet (@pascal_bornet) 17 décembre 2024
Safety equipment isn't just about protection anymore—it's about enhancing human potential through smart technology.
When innovation meets safety, we don't just prevent accidents—we create possibilities.
🤔… pic.twitter.com/uuEby7oOzFInnovation Steps Into Safety: When Technology Protects Every Step Safety equipment isn't just about protection anymore—it's about enhancing human potential through smart technology. When innovation meets safety, we don't just prevent accidents—we create possibilities.
-
India AI Mission Deepfake Detection Tools Safety Initiative
By
–
Develop tools to identify and verify the authenticity of media, safeguarding society from the harms of deepfakes. Click here to know more: https://
indiaai.gov.in/article/expres
sion-of-interest-for-safe-trusted-ai-projects-under-indiaai-mission
… @GoI_MeitY @abhish18 @_DigitalIndia #deepfakes #AIsafety #authenticity #usersafety -
Meta Opens AI Models for U.S. Military Use
By
–
Meta Opens Its AI Models for the (U.S.) Military https://
bit.ly/3CZK6ly